Novel Machine Learning and Design Methods to Improve EEG-based Motor Imagery Brain-Computer Interfaces

Brain-computer interface (BCI) systems read and infer brain activity directly from the brain through for instance electroencephalography (EEG), while bypassing the common neuromuscular pathways. These systems can provide assistive devices for communication and locomotion, interventions for motor restoration and rehabilitation, as well as neurofeedback training for cognitive enhancement. Motor imagery-based BCIs are a category of BCI systems that allow a user to generate control commands by imagining moving different body parts (such as the right or left hand) and the goal of the BCI is to correctly detect the imagined body part from EEG patterns. The ability of users to generate discriminable motor imagery-based control signals is very variable and environmental effects such as other brain processes affect current BCI systems to the point that they are mostly limited to lab environments. This dissertation proposes novel machine learning and design methods to improve the performance and reliability of motor imagery EEG-based brain-computer interfaces to provide patients in need, the ability of independent interaction with the outside world. This dissertation proposes: 1) a more elaborate feedback paradigm to allow users to better learn how to generate discriminable motor imagery signal, 2) a novel hybrid motor imagery BCI utilizing the user's brain activity in response to the BCI output (feedback) with improved performance and reliability compared to existing motor imagery BCI systems, 3) an artificial neural network architecture to capture spatio-temporal aspects of the motor imagery signal to improve classification of motor imagery activity, and 4) a novel covariance-based method that uses Riemannian and Euclidean geometry to capture spatial and temporal aspects of the feedback-related brain activity in response to BCI error to further improve hybrid BCIs that utilize this type of brain activity . By giving improved training feedback and better utilizing the available brain signals, the proposed methods improve the performance and reliability of motor imagery BCIs and provide the chance to greatly increase the number of people who are successfully able to operate one. The developed techniques could also be useful for discovering and training other mental commands that could be used in EEG-based BCIs not limited to motor imagery BCIs